Speaker

Biography

Abstract 

Stephen Burgess, PhD 
Group leader
University of Cambridge

Stephen completed his BA and MMath (Part III) in Mathematics at the University of Cambridge. He studied for a PhD in the MRC Biostatistics Unit, Cambridge, from 2008-11 working on methods for Mendelian randomization analysis under the supervision of Simon Thompson. He joined the Cardiovascular Epidemiology Unit in the Department of Public Health and Primary Care of the University of Cambridge in 2011. In 2013, Stephen was awarded a Wellcome Trust entry-level fellowship (Sir Henry Wellcome Post-doctoral Fellowship) to continue theoretical and applied work in the field of Mendelian randomization. In 2017, he moved to the MRC Biostatistics Unit on a Wellcome Trust/Royal Society intermediate fellowship (Sir Henry Dale Fellowship) to establish a research group in the MRC Biostatistics Unit, where he is now a Programme Leader. In 2023, he was given a Career Development Award by the Wellcome Trust to continue work in this area. He leads a small team of researchers split between the MRC Biostatistics Unit and the Cardiovascular Epidemiology Unit. He is always open to requests for collaboration, either on theoretical or applied Mendelian randomization projects.

Instrumental variables in observational data: what are they, and what do they allow us to estimate?

A natural experiment is an observational study that enables causal conclusions to be drawn. In a natural experiment, individuals are divided either at random or in a way that mimics randomization (known as quasi-randomization), but this division is performed by a natural process or an artificial distinction rather than by an investigator. The variable that divides individuals is known as an instrumental variable.

In this talk, I will provide an introduction to instrumental variables: what they are, and how they can be used to make causal claims. I will then present Mendelian randomization, the use of genetic variants as instrumental variables. Due to inherent randomness in the process of genetic inheritance, genetic variants act somewhat like randomization. As the majority of drugs influence proteins and genes encode proteins, there is a specific relevance for Mendelian randomization to the drug development pipeline. Mendelian randomization can provide evidence on target validation, re-purposing, safety signals, mechanistic relevance, and effect heterogeneity: are pathways worth drugging, what outcomes do they affect, how do they affect them, and who would most benefit from intervention? The talk will be illustrated with examples of targets for existing and emerging drugs.

Finally, I will talk about the role of estimation in natural experiments, and in Mendelian randomization in particular. A distinction will be drawn between investigations where the aim is estimation versus those where the aim is aetiology (i.e. is the risk factor a cause of the outcome?). In a natural experiment, while it is valuable to define the estimand carefully, the investigator cannot control the intervention, and so has limited choice of what parameter to estimate.

Jack Bowden, PhD 
Professor of Biomedical Data Science and Senior Director
Novo Nordisk and University of Exeter

Jack Bowden is an internationally recognised biostatistician, with 20 years’ experience across academia (at the Universities of Cambridge, Bristol and Exeter) and the pharmaceutical industry (at Amgen, Novartis & Novo Nordisk).

In his dual position as Professor of Biomedical Data Science at the University of Exeter and Senior Scientific Director at Novo Nordisk, he continues to lead teams of researchers in the development statistical tools and methods to enable impactful scientific research.  Key areas of statistical expertise include clinical trials, genetics and omics, epidemiology, real world data, evidence synthesis and meta-analysis, with bias adjusted estimation and causal inference being unifying themes.  He has been a Clarivate Highly Cited Researcher each year since 2019 in recognition of his research standing (with the caveat that this is biased in favour of men).

Instrumental variable methods in randomized trials: why do we need them and what do they offer?

In the analysis of RCTs, randomized assignment is thought to facilitate such simple and direct inference about the effect of a treatment, that the word `causal’ is deemed redundant and is almost never used. To the average trialist, ‘causal inference’ conjures images of complex analysis on messy observational data, enabling the user to make outlandish claims based on opaque, untestable assumptions. So much so, the word `causal’ has been banned from observational analysis publications altogether by several journals.

In this talk I will describe why RCTs can have more in common with observational studies than is generally realised, why it is useful to define causal effects explicitly in RCTs, how these effects differ from what is targeted in a typical trial analysis, and how they can be estimated simply and transparently using Instrumental Variables (IVs). I will use the STEP 1 randomized trial as an exemplar study: it evaluated the effect of taking semaglutide versus placebo on body weight over a 68-week duration, but as with any study evaluating an intervention delivered over a sustained period, some nonadherence was observed. I will describe a recent IV method we developed to adjust for this, which allows a patient’s complete adherence history to influence their outcome at a given time point and facilitates estimation of a hypothetical causal Estimand: the treatment effect if all trial participants would have adhered to both semaglutide and placebo.